Hugging Face's logo

jarrelscy
/
GLM-5.3-Vision-NVFP4-ARVQ-v2-hybrid

Safetensors
vllm
glm5v
arvq
nvfp4
experimental
8-bit precision
nvfp4_arvq_hybrid
Model card Files
xet
 Community
GLM-5.3-Vision-NVFP4-ARVQ-v2-hybrid / METHODOLOGY.md
jarrelscy's picture
jarrelscy
Publish quantization/data reproduction source and correct recipe documentation
cb95231 verified
preview code
|
Raw History Blame Contribute Delete
917 Bytes

ARVQ-v2 methodology correction

The previous file was copied from v1 and described the wrong target, scale dtype, corpus and schedule. The current methodology is documented in reproduce/README.md, with exact stage order in RUNBOOK.md and settings in recipe.json.

Key facts: same-input target; original FP8 expert teacher; FP4 per-expert books; FP16 block scales;15,007,754-token cleaned corpus;50× next-boundary loss weighting; reassignment every10; decay by25 or earlier; patience10. Validation selection is unweighted target-relative L2. Most layers stop early. All75 layer receipts remain available. Historical contradictory metadata is archived under reproduce/historical_metadata.

The bundled reference packer/decoder now handles both uint8 E4M3 and actual FP16 scales, tested on CPU. No model weights were changed by this correction.